Artificial Intelligence Enabled IoT-based System for Environmental Monitoring: Design and Evaluation

Advances in Internet of Things (IoT) technologies have presented new epoch in environmental monitoring, leading to the deployment of billions of sensor devices in various domains to sense and monitor the environment.
IoT sensors has the capacity to provide high resolution spatio-temporal dataset in environmental monitoring. The use of Low-Cost IoT Sensors (LCS), however, will provide wider scope and adoption for this purpose. LCS are imperfect and are prone to recruiting erroneous data and their application in environmental monitoring raises some important questions, especially pertaining to the reliability, in-field performance and data quality of the sensors. IoT technologies are still at the early stage of development in many application areas including in the area of ecological sensing and as such they are challenged by several issues including data handling and framework. There is currently no framework or formalized architecture in existence for the application of LCS devices in environmental monitoring and no complete tool has been developed for data handling and analysis. The objective of this thesis is to contribute to research efforts on improving field performance and data quality of LCS devices through the development of appropriate data processing and data modelling tools. The work leverages the capabilities of AI and Machine learning (ML) to support the effective application of IoT in environmental monitoring. In particular, the research aim to 1.) Develop data acquisition model that employs low-cost IoT sensors to capture large scale and longitudinal observations, 2.) develop specific data handling and Data Quality Improvement (DQI) technique for LCS applied to the peatland monitoring domain; 3.) develop a strategy based on efficient data imputation to support on-site sensor calibration; 4.) implement novel and efficient method for proxy-estimation of pollutants, to quantify the concentration of a gas for which no specific sensor is included in a multi sensor device and 5.) Present high fidelity and robust data on peatlands’ carbon stocks, water table depth and soil moisture content. Five major contributions are made in this thesis, first, a pervasive framework for ecological sensing is proposed, to facilitate the integration of LCS devices into current monitoring solutions. The framework details procedures for data handling and analysis of LCS data. In a second step, the challenges facing IoT sensors are investigated to particularly understand how sensor behavior is impacted by environmental conditions. This led to the design of sensor calibration model which is able to correct the effects of meteorological conditions such as Temperature (T), Relative humidity (RH) on LCS responses. The third contribution focuses on evaluating different data imputation techniques to better understand their capability in handling data sparsity and missing values issues for LCS and further shows how imputation impacts sensor calibration. Consequently, the fourth contribution proposes ProxySense; a novel technique to estimate the concentration of a pollutant gas for which no specific sensor is included in a multi-sensor array. Finally, the tools and techniques developed within this project is applied to a real world monitoring site. The Irish peatlands was chosen as the practical context for this research and over three years instrumentation of the site have enabled the collection of large dataset on CO2 carbon stock, and soil moisture contents which is freely available and can be assessed by other researchers and stakeholders.